AI@HZDR in the Research Field MATTER
In HZDR's largest research area, scientists use AI methods to analyze large amounts of data or accelerate simulations. For example, researchers at the Center for Advanced Systems Understanding (CASUS) develop machine learning algorithms and methods, as well as software for the world’s largest exascale computers such as the FRONTIER supercomputer in the United States. These advances enable highly versatile and ultra-fast simulations. Remarkably, many of the new AI models can be be adapted to solve a wide range of problems.
Future Materials Thanks to Accelerated Simulations
Snapshot of a deep learning simulation | Image: HZDR/CASUS
Calculating the properties of every conceivable combination of materials on the computer instead of conducting time-consuming experiments in the laboratory – new information and communication technologies are likely to benefit from this just as much as the energy transition or drug development. This is made possible by the MALA (Materials Learning Algorithms) software package, which combines deep learning methods (see glossary) with physical approaches. The starting point for the machine learning model, which was developed at the HZDR institute CASUS, is the atoms arranged in space with their respective neighbors. Trained accordingly, MALA can predict the electronic structure of a material – for more than a hundred thousand atoms and at unprecedented speed.
The MALA approach is particularly suitable for high-performance computing (HPC). With increasing system size, MALA enables independent processing on the computing grid it uses, which is why the software package is characterized by extremely efficient handling of HPC resources – especially graphics processors.
In addition to the development partners HZDR and Sandia National Laboratories, MALA is already being used by institutions and companies such as the Georgia Institute of Technology, North Carolina A&T State University, Sambanova Systems Inc., and Nvidia Corp.
Contact:
Dr. Attila Cangi
Head of Department: Machine Learning for Materials Design
More information:
►Press release: Prestigious US prize awarded to CASUS researchers
►Press release: Machine learning takes materials modeling into new era
Electromagnetic fields excite oscillations in a magnetic vortex. The non-linear interaction is similar to the interplay between neurons and synapses in the brain and can be used for pattern recognition. | Image: H.Schultheiß/HZDR
Brain-inspired Computing
In future, autonomous driving systems will have to record data from distance, motion or speed sensors at lightning speed and evaluate it in exact chronological order. Currently, this work is done by software that runs on conventional computer systems. Although specially developed neural networks can process such fast sensor data, they are usually overwhelmed by the task of doing this in real time and in the correct temporal sequence due to the conventional computer architecture.
Researchers at the HZDR Institute of Ion Beam Physics and Materials Research want to solve this problem together with international colleagues and chip manufacturers and in the EU-funded NIMFEIA project.
Their new hardware is based on micrometer-sized magnetic disks. If these are excited with electrical pulses in the microwave range, waves are created in the magnet. The quantum particles of these waves, the magnons, can decay into many more magnons. This allows patterns to be recognized over time and processed in real time. The processes involved are similar to the signal storage and processing of networked neurons in the brain – right down to the fact that neurons only transmit the stimuli they receive if they are above a certain threshold value. This also applies to the magnons, which have to overcome an activation threshold. Another advantage is that neuromorphic technology requires only a fraction of the energy that conventional systems need.
Contact:
Dres Helmut and Kathrin Schultheiß
Research group Spin: Interaction and Control
More information:
►Press release: Dancing Magnons
Digital Twin Improves Experiments on High-power Lasers
Image from the vacuum chamber of the DRACO laser plasma accelerator | Image: HZDR/André Wirsig
When the flash of light from a high-power laser hits matter, a vast number of processes take place almost simultaneously on different size scales. Researchers can neither measure these directly in experiments nor describe them comprehensively in theory. This is why data scientists are working with physicists from the HZDR Institute of Radiation Physics to develop extremely powerful computer simulations and digital twins of the experiments. They use the versatile plasma simulation code PIConGPU as a basis. If, for example, it is possible to reconstruct the path of the particles as accurately as possible, the DRACO and PENELOPE laser systems can be driven to ever higher intensities and the quality of the proton pulses generated can be improved.
Contact:
Prof. Ulrich Schramm
Director Institute for Radiation Physics and Head Laser Particle Acceleration Division
Dr. Michael Bussmann
Deputy Director CASUS – Center for Advanced Systems Understanding
Head Computational Radiation Physics
More information:
►Website: Laser-plasma-acceleration for ultra-high dose rate radiobiology
Faster Path to Synthetic Data
Surrogate models of neural networks in high-energy physics | Image: CASUS/2021 CMS Collaboration
In fundamental physics research, experimental data is often supplemented by synthetically generated data. With currently available simulation methods, this is a time-consuming undertaking that ties up immense computer capacities. A cooperation project between DESY – Deutsches Elektronen-Synchrotron and HZDR is dedicated to an approach with which data on the behavior of physical systems can be generated more quickly using neural networks.
The aim of the project “SynRap - Machine learning-based generation of synthetic data for rapid modeling in physics” is to accelerate the process by a factor of 1,000. To this end, the DESY scientists want to compile a toolbox of suitable machine learning algorithms together with their CASUS colleagues from the HZDR. These algorithms originate from a specific subgroup of neural networks – in contrast to deep neural networks, these are so-called surrogate neural networks or surrogate models of neural networks.
The special feature of the project is that the toolbox with various neural networks is to be used in many areas of research.
Contact:
Dr. Attila Cangi
Head of Department: Machine Learning for Materials Design
More information:
►Press release: Helmholtz Association funds project for data acquisition using neural networks
Predicting Extreme Conditions with AI
The most extreme conditions prevail inside planets | Animation: HZDR/Science Communication Lab
NASA's Kepler space telescope has detected thousands of exoplanets in distant galaxies. Experts can determine the size of such a planet, whether the temperature on its surface is suitable for life and what chemical elements it is made of by observing it. Many questions – for example about its internal structure or the existence of a magnetic field – remain unanswered. Data scientists at the HZDR Institute CASUS in Görlitz are currently developing a digital twin that will contribute to a better understanding of exoplanets based on the available data and different probability assumptions. In this way, they hope to close important gaps in our knowledge of alien planetary systems. In addition, a new simulation method enables reliable predictions for future experiments for the first time – for example at the Helmholtz International Beamline for Extreme Fields at the European XFEL.
Contact:
Dr. Attila Cangi
Head of Department: Machine Learning for Materials Design
Dr. Tobias Dornheim
Young Investigator Group Leader: Frontiers of Computational Quantum Many-Body Theory
More information:
►Press release: Rolling dice to gain insights into planets and stars